Unsupervised Channel Compression Methods in Motor Prostheses Design.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• High-performance brain-machine interfaces (BMIs) require extensive neural recordings.
• Current implantable neural interfaces face power limitations due to linear scaling with recording channels.
• Reducing power consumption is critical for advancing implantable neural technologies.

Purpose of the Study:

• To develop and evaluate an unsupervised-learning-based compressed sensing strategy for neural data.
• To design novel neural interface architectures that compress neural data by combining spiking activity channels.
• To minimize information loss in latent variables during neural data compression.

Main Methods:

• Developed an entropy-based compression strategy modeling neural populations from lower-dimensional latent variables.
• Evaluated compressed features by inferring latent variables and estimating held-out neural activity and arm movements.
• Applied methods to cortical regions (PMd and M1) and compared against random projections and supervised channel dropping.

Main Results:

• The unsupervised learning approach effectively compressed neural data while preserving essential information.
• Accurate estimation of held-out neural activity and arm movements was achieved using compressed features.
• The proposed method demonstrated competitive or superior performance compared to traditional strategies.

Conclusions:

• Unsupervised learning-based compressed sensing offers a viable strategy for reducing power requirements in BMIs.
• This approach facilitates the development of more efficient and scalable neural interfaces.
• The method holds promise for advancing the capabilities of future brain-machine interfaces.